Uniform-Baseline Implementation Matters in Long-Video Question Answering: A Controlled Study of Frame Sampling

Abstract

Long-video question answering (QA) under a tight frame budget forces a sampling choice, and query-aware keyframe samplers are assumed to outperform uniform sampling. We investigate which properties of a frame selection are associated with a vision-language model’s (VLM) accuracy under a fixed protocol. We compare four samplers — deterministic uniform sampling (Uniform-strict), two published query-aware methods (BOLT and Adaptive Keyframe Sampling (AKS)), and a signal-free bucket-random control — on one backbone (Qwen2.5-VL-7B), one 16-frame budget, and one benchmark (the Video Multi-Modal Evaluation (Video-MME) long subset); because BOLT additionally forwards frame-rate metadata, it is only a supporting comparison; the controlled contrast is Uniform-strict versus bucket-random. Under this protocol, we make two observations: the choice of uniform-baseline implementation can materially affect the measured gap, and matched segment coverage does not imply matched accuracy — the bucket-random control preserves coverage yet loses 9.5 percentage points (pp). A median gap coefficient of variation (CV) of 0.007/0.279/0.400/1.830 aligns with accuracies of 61.5%, 54.0%, 52.0±1.8% (mean ± standard deviation [SD], five seeds), and 51.5%, identifying frame-interval regularity as a diagnostic correlate of accuracy and motivating controlled tests of its causal role.